accelerate
https://github.com/huggingface/accelerate
Python
🚀 A simple way to train and use PyTorch models with multi-GPU, TPU, mixed-precision
Triage Issues!
When you volunteer to triage issues, you'll receive an email each day with a link to an open issue that needs help in this project. You'll also receive instructions on how to triage issues.
Triage Docs!
Receive a documented method or class from your favorite GitHub repos in your inbox every day. If you're really pro, receive undocumented methods or classes and supercharge your commit history.
Python not yet supported1 Subscribers
Add a CodeTriage badge to accelerate
Help out
- Issues
- Add an opt-in fp32 upcast of low-precision trainable params under mixed precision (DDP, single device), and warn by default
- Fix TP + FSDP2 prepare without the removed ReplicateParallel, refuse cpu_ram_efficient_loading with TP, add numerical parity tests
- Fix tied parameter matching for repeated module names
- Bind each model chunk's index in MegatronEngine param_sync_func
- Preserve distinct shared-storage views in safetensors checkpoints
- Fix `gather_for_metrics` remainder for sized `IterableDataset` with `dispatch_batches=False`
- Fix disk offload for module extra state
- Fix load_kwargs for PyTorch model checkpoints
- Quick survey: your experience as a PyTorch Ecosystem maintainer
- Keep a split module's own buffers in infer_auto_device_map
- Docs
- Python not yet supported